Expressive Loops

.gitignore

@@ -1,8 +1,16 @@
 /docs/public
 /tests/test_real_time/
 /tests/tests_output
-/models
-.MMM/
+runs/models
 .idea
 __pycache__
 **/__pycache__
+/.MMM
+/debug
+/tests/output
+/tests/test_midis
+
+logs/*
+CCLOG/*
+models/*
+runs/*

README.md

@@ -88,6 +88,12 @@ python scripts/train_tokenizer.py
 python scripts/train_model.py
 ```
 
+## Pretrained model for inference
+Download the model here:
+https://drive.google.com/file/d/1Up6OxUANSUcBipPz_V9hLj6EiFTPCmiO/view?usp=sharing
+
+Use generate.py to generate samples using the mentioned model.
+
 ## Data preprocessing
 
 1. Filter non-valid files: corrupted or less than 8 bars;

docs/examples.rst

@@ -67,7 +67,7 @@ Generate
     # Instantiate the tokenizer
     MMM(params="/path/to/tokenizer.json")
 
-    output_scores = generate(
+    output_scores, metadata = generate(
             model,
             tokenizer,
             inference_config,

docs/training.rst

@@ -10,8 +10,19 @@ The only thing needed is to use the provided :class:`mmm.DatasetMMM` dataset cla
 1. Loads the music file;
 2. Shuffles the order of the tracks, and randomly keeps a certain number of tracks (``ratio_random_tracks_range``);
 3. Performs data augmentation by randomly shifting the pitches, velocities and duration values;
+    Currently:
+        +- 6 pitch offset;
+
+        +- 2 Velocity offset;
+
+        0 duration offset;
 4. Select randomly whether to perform bar infilling or to generate a new track (``bar_fill_ratio``);
-5. Reduce the sequence length so that it does not exceed the limit (``max_seq_len``);
+    Currently, bar_fill_ratio = 0.75
+5. If infilling, select a number of bars to infillin equal to (``tot_bars * n_bars_infilling_ratio``)
+    where ``n_bars_infilling_ratio`` is random from 0.1 to 0.4
+6. Randomly select a set of attribute controls which are going to be computed on a random subset of the bars to be infilled (if infilling).
+7. Reduce the sequence length so that it does not exceed the limit (``max_seq_len``)
+    where ``max_seq_len`` = 2048
 6. If bar infilling: randomly selects the portion supposed to be infilled and move the associated tokens at the end of the sequence for the model to learn;
 
 The DatasetMMM object

mmm/__init__.py

@@ -2,7 +2,7 @@
 
 from .config import InferenceConfig
 from .data_loading import DatasetMMM
-from .inference import generate
-from .logits_processor import StopLogitsProcessor
+from .inference import generate, generate_batch
+from .logits_processor import InfillLogitsProcessor, TrackLogitsProcessor
 
-__all__ = ["DatasetMMM", "generate", "InferenceConfig", "StopLogitsProcessor"]
+__all__ = ["DatasetMMM", "generate", "generate_batch", "InferenceConfig", "InfillLogitsProcessor", "TrackLogitsProcessor"]

mmm/config.py

@@ -21,9 +21,9 @@ class InferenceConfig:
         the new tracks
     """
 
-    bars_to_generate: dict[int, list[tuple[int, int, list[str]]]] | None = None
-    new_tracks: list[tuple[int, list[str]]] | None = None
     context_length: int = 4
+    bars_to_generate: dict[int, list[tuple[int, int, list[str], str]]] | None = None
+    new_tracks: list[tuple[int, list[str]]] | None = None
     autoregressive: bool = False
     infilling: bool = False